Late chunking · Deferred splitting
Late chunking is a variant of chunking where the embedding is computed on the entire document before splitting, which preserves each passage’s context. It was described by Jina AI in 2024 with long-context embedding models.
Standalone definition · 35 words
How does late chunking work?
In the classic approach, you split first and then encode each piece in isolation, which loses context: a “he” or “this company” becomes ambiguous. With late chunking, the model encodes the whole text at once thanks to a long context, then vectors are grouped by passage. Each chunk keeps information from the entire document.
What are the benefits and limits?
It improves retrieval when passages depend on context, at the cost of a long-context model and heavier computation. It concerns search system designers, not site publishers: no mainstream engine has documented using it.
Editorial consequence
Even if a system uses late chunking, a standalone passage remains the best bet: it works with classic chunking and with contextual variants. Naming the subject in each section avoids relying on implicit context.
Frequently asked questions
Should I adapt my content to late chunking?
No, nothing specific. Good standalone-passage practices are enough.
Are late chunking and contextual chunking the same?
They are two approaches to reintroducing context into chunks, with different methods.
Who uses it?
RAG systems built on long-context embedding models.
Related terms
Further reading
- Checklist GEO and LLM / AI 2026
- 2026 GEO Tools Comparison: The Decision Guide
- ← Back to the GEO glossary (114 terms)